Uplift

Uplift: surviving a tt-metal / tt-umd pin bump

When the tt-metal pin moves forward (e.g. a companion change merges to main, or you re-pin to a newer commit), you import two moving parts at once: the tt-metal source and the tt-umd submodule it points at. Either can regress the emule build because emule substitutes its own runtime (SWEmuleChip, emulated_program_runner.cpp) and JIT kernel shims (include/jit_hw/) for the real hardware path. This skill is the methodology for turning a red post-bump regression into a proven root cause and a minimal, faithful fix.

The governing principle, learned the hard way: prove it by building and running, not by reading git log. Pickaxe/git show/blame routinely mislead (moved code, squash merges, shallow submodule object DBs that render a commit as a whole-tree add). A bisection that compiles and runs an oracle is the only authority.

When to invoke

• A pin bump (or tt-metal-pin.txt change) makes a previously-green regression fail.
• Symptoms after a bump: device-open aborts (SIGABRT/SIGSEGV during start_device), jit_compile_kernel: compiler failed (use of undeclared identifier ...), a deterministic hang, or new ATOL/PCC mismatches across many unrelated ops.
• You suspect the bundled umd bump rather than tt-metal source, and need to isolate which.

For a single missing compute shim (one <op>_tile), use /compute-llk-bringup. For data-corruption on a kernel that runs without crashing, use /memory-debug. This skill is for broad regressions that a pin bump introduced.

Step 0 — Read the failure correctly, don't trust the label

• Inspect artifacts, not the check's green/red. A C++ regression job can show "pass" while its result artifact shows aborts — known-failure allowlists and classify steps can absorb failures, and exit codes don't always propagate. Open the actual run log/artifact and read the per-test verdicts.
• Classify the failure mode before chasing — they have different root-cause paths:
  • device-open crash (fails in the device fixture / start_device, before any op) → a shared runtime/reset/topology path, almost always umd or the emule chip.
  • JIT-compile error (undeclared identifier, header not found) → a jit_hw shim drift; the kernel can't compile against emule's headers.
  • hang → often a JIT-infra issue (e.g. forking a compiler from a many-threaded process), not the op.
  • data mismatch → op-correctness; usually /memory-debug territory.
• Per-arch scope. Supported arches gate the decision; an out-of-scope arch is typically allowlist-gated and expected-broken — don't let its churn block a bump, but don't silently regenerate its allowlist over a real new regression either.

Step 1 — Build a fast, deterministic oracle

Pick the smallest, fastest test that reproduces the failure with an unambiguous signal, and wrap it so it emits exactly GOOD / BAD / SKIP:

• GOOD = clean pass.
• BAD = the specific signature (e.g. a particular abort message + address, or the exact compiler error). Match the signature, not just "non-zero exit", so an unrelated failure can't masquerade as the bug.
• SKIP = build break or any other outcome → maps to git bisect 125.

Step 2 — Bisect: isolate metal-source vs the umd bump

The pin range moves both metal and the umd gitlink, so bisect in two stages.

1. Metal bisect. git bisect start <new-pin> <old-pin> in tt-metal, then git bisect run a script that, per commit: updates the umd submodule to that commit's recorded gitlink, builds only the oracle's target (not the whole tree), runs the oracle, and returns 0/1/125.
   • If the first-bad commit changes only the umd gitlink, the regression rides with the umd bump — proceed to the umd bisect.
   • If it's a metal-source commit, the cause is in tt-metal; triage that commit.
2. UMD bisect. Pin metal at the first-bad (umd-bump) commit, then git bisect start <bad-umd> <good-umd> inside the submodule and bisect with the same oracle, varying only umd. Build breaks on old-umd-vs-newer-metal become SKIP.

Bisection rules that keep it honest:

• Hold the emule layer fixed (the jit_hw headers and any local emule source) for the whole bisect — the variable under test is metal+umd, exactly what the bump changed.
• Build the minimal target so each step is a fast incremental relink.
• Automate with git bisect run so you don't hand-judge steps.
• Restore all repos (git bisect reset, re-checkout the intended pin, resync the submodule) when done.

Step 3 — Prove the boundary and the mechanism

• Controlled before/after: build+run the oracle at culprit^ (GOOD) and culprit (BAD) with everything else fixed. That's the proof.
• Read the mechanism from a two-tree diff, not git show. git show <culprit> can render a whole-tree add in a shallow/grafted submodule; use git diff <culprit^> <culprit> -- <file> between two checked-out trees.
• Read the upstream PR the culprit came from. It is the authoritative spec for any emule-side mirror — its diff tells you the exact old→new API or behavior mapping to replicate.

Step 4 — Design the fix (two recurring classes)

Most uplift regressions are one of these.

A. Behavioral regression in a shared runtime path

Upstream changed a path that emule used to neutralize. Example shape: a device bring-up step that used to dispatch through a per-chip virtual the SW chip overrode as a no-op now issues an unconditional register write that the emule memory model can't satisfy.

Fix principles:

• Preempt at the right layer, using an existing mode/chip-type signal the code already branches on. Gate strictly on the emulated case so silicon / simulation / other modes are untouched.
• No magic constants when a canonical accessor exists; prefer "skip the operation for this mode" over "match a hardcoded address".

B. API-surface drift in the JIT kernel shims (jit_hw)

Upstream renamed/refactored a kernel-facing API (an enum collapse, a template signature change, a moved header) and the emule shim still exposes the old surface, so kernels fail to JIT-compile with undeclared identifier.

Fix principles:

• Mirror the upstream change exactly in the shim — same enum/struct/signature surface — using the upstream PR diff as the spec.
• Preserve emule's bodies (host memcpy / no-op / emulated sync). Only the option/signature surface changes; the emulated semantics stay.
• Migrate all dependents in the same surface (the header's call sites, sibling headers that reference the renamed symbols) so nothing dangles. Leave genuinely independent/older surfaces (e.g. a separate experimental/ API with its own self-contained definitions) alone.
• Keep .claude/references/structure.yaml in sync when symbols are added/removed.

Step 5 — Wire the fix through the repo chain

A fix can span tt-umd → tt-metal → tt-emule. Order and push mechanics matter:

• Push dependencies before dependents. If the metal commit bumps the umd gitlink to a fix commit, that umd commit must be on its remote first, or the submodule pointer dangles for CI.
• Never push without explicit go-ahead; for tt-emule's own PR, bump tt-metal-pin.txt to the (pushed) metal fix commit, using the full SHA.
• Re-sync the vendored e2e demo on any pin bump. The end-to-end model lane (docs/e2e-models.md) drives an emule-vendored copy of simple_text_demo.py at tt-metal/tests/emule/models/test_tt_transformers_text_demo.py, which tracks upstream only as of the pin SHA in its header. When you move tt-metal-pin.txt, re-run git show <new-pin>:models/tt_transformers/demo/simple_text_demo.py into the vendored file and re-apply its two # emule: markers. The demo is not in the C++ regression, so this drift is otherwise uncaught until the nightly e2e lane runs against the new pin.
• jit_hw edits need no C++ rebuild. Shims are compiled at JIT time, so a jit_hw change takes effect on the next kernel compile — clear the JIT cache and re-run; iteration is fast.

Step 6 — Verify recovery and prove no regression

• Oracle before/after on the built fix: the BAD signature is gone.
• Broad sweep: run one test per previously-failing file. The headline invariant is that the crash-signature count goes to zero; PASS count rises with no new failures.
• Triage every residual into exactly one bucket: (a) caused by the fix — must be zero; (b) the original bug — must be zero; (c) a separate pre-existing gap that was merely unmasked once earlier crashes stopped. A fix that doesn't touch op math can't cause a data mismatch — reason from that.
• Don't conflate a sweep with the curated suite. A one-test-per-file sweep runs the first collected test in each file, which is usually not what the curated regression script selects (it uses -k/::node subsets). Before treating a sweep failure as real, check whether the curated script even exercises that node:
  • Map the failing node to the script's exact run_pytest args and resolve membership with --collect-only.
  • Validate your collection-parsing first. Verbose pytest config (-vvs) makes --collect-only print a <Function ...> tree, not file::node lines — grep for the <Function ...> reprs, and confirm your grep matches a test you know is selected before trusting any zero counts. (Class-based tests are Class::method; a bare file::method nodeid will error.)

Step 7 — De-scope and track what isn't yours

• A residual that is a separate pre-existing gap (e.g. a missing compute-LLK shim, a known SIGFPE, an op-correctness mismatch) should be filed and tracked, not used to block the uplift, and not "fixed" by widening the uplift fix's scope.
• If the uplift's root cause is upstream-intended (a legitimate refactor), the emule side is what must adapt — the fix belongs in the emule layer (umd SW chip / jit_hw), not a revert.
• Note any cross-arch implication (a shim/runtime fix for the supported arch often also addresses the same root cause flagged for an out-of-scope arch) but don't chase the out-of-scope arch in the same change.

Anti-patterns (consolidated)

1. Concluding from git log -S / git show / blame. They mislead on moved code, squashes, and shallow submodules. Bisect (build+run) is authoritative.
2. Trusting a green CI label. Read the artifact; allowlists hide aborts.
3. A range-based guard that swallows real errors. Gate on the specific case/mode, not "anything that would fault".
4. Intercepting low-level I/O to dodge a high-level behavior change. Preempt the operation at the layer that owns it.
5. Hardcoding magic values when a canonical accessor or a mode flag exists.
6. Pushing a dependent before its dependency (metal gitlink before the umd commit it references).
7. Reading a one-test-per-file sweep as the curated suite. Check --collect-only membership, with a validated parser.

References

• /compute-llk-bringup — for a single missing compute <op>_tile shim (the common Class-B fix when the drift is one op).
• /memory-debug — for data-corruption residuals (Class-C, op math).
• docs/api-injection-points.md — where emule intercepts the pipeline (helps choose the "right layer" in Step 4A).
• references/emule-mapping.md — HW concept → emule strategy vocabulary.
• BUILD_GUIDE.md — build/test setup, the targets to build for an oracle, and the regression scripts.